Estimation of the convolutional neural network with attention mechanism and transfer learning on wood knot defect classification

Abstract

In the intelligent production process of wood products, the classification system of wood knot defects is a very practical solution. However, traditional image processing methods cannot handle it well due to the uncertainty of manually extracted features. Therefore, a lightweight and reliable artificial neural network model is proposed to classify and identify our objective. To solve this problem, a wood knot defect recognition model named SE-ResNet18 combining convolutional neural network, attention mechanism, and transfer learning is proposed in this paper. First, the Sequence-and-Exception (SE) module is combined with Basicblock and is constructed as two modules called RBBSE-1 and RBBSE-2. These modules learn to enhance features that are useful for the current task, suppress useless features, and fuse the output features with the original features. Then, the fully connected layer is replaced with a global average pooling layer, which can effectively reduce the parameters of the fully connected layer in the model. Finally, a SE-ResNet18 was constructed by one convolutional layer, five RBBSE-1 modules, and three RBBSE-2 modules of different channels. The SE-ResNet18 has a higher accuracy (98.85%) in the test set compared to the unimproved model ResNet-18. Compared with the previously proposed ReSENet-18, more SE modules are used in SE-ResNet18 to provide a basis for future training on a larger-scale dataset. Based on the same test set, a comparison with other classical models (such as LeNet-5, AlexNet, etc.) was conducted, and the results validated the superiority of the proposed model. The proposed model achieves the expected objective and provides a new way of thinking for non-destructive testing of wood.